Spiking Neural Network (SNN) With Memristor Synapses Having Non-linear Weight Update

Area of Science:

• Neuromorphic Engineering
• Artificial Intelligence
• Materials Science

Background:

• Spike Neural Networks (SNNs) mimic brain's energy-efficient signaling.
• Memristors are promising for SNN hardware but have non-ideal properties.
• Device non-idealities challenge SNN implementation.

Purpose of the Study:

• To investigate the impact of non-linear weight updates in memristor-based SNNs.
• To determine conditions for maintaining SNN performance despite device non-idealities.

Main Methods:

• Simulated SNNs incorporating a memristor device model with non-linear weight updates.
• Analyzed network performance and accuracy under varying device non-linearity conditions.

Main Results:

• SNNs exhibit strong tolerance to device non-linearity.
• High network accuracy is maintained if LTP/LTD curves are symmetric or non-linearity factors are positive.
• Analysis revealed balance in network parameters and weight variability contribute to tolerance.

Conclusions:

• Memristor-based SNNs can achieve high accuracy despite device non-idealities.
• Specific conditions on weight update linearity ensure robust neuromorphic hardware.
• Findings provide guidance for designing future emerging device-based neuromorphic systems.